Enable full state runs with scaling about covasim

[inc0]

Right, this is mainly to produce numbers per day for big populations, probably good to document when not to use dynamic resampling

[ckerr-IDM]

We won't want to start from a single infection -- too much stochasticity anyway. This question is a good one, related to this:
https://github.com/amath-idm/covasim/issues/23

[inc0]

Idea @cliffckerr had and I'll prototype:

When we reach certain threshold of infections (say, 50%, but configurable) we scale up frame of reference by 2 (or later, configurable).
That means:

1. De-infect random 50% of infected
2. Resurrect 50% of dead
3. Re-susceptible(?) 50% of recovered
4. Multiply every summary stat by 2

[cliffckerr]

@inc0 sounds good. the main change that needs to happen is to expand the definition of scale to have it be an array of length n_days rather than a scalar (or better, keep scale a scalar but introduce a scalevec parameter for this). but yeah basically just loop over 50% of all non-susceptibles and restore them to the susceptible state (no method for this currently, but should be added)

[dlchao]

Doesn't scaling affect elimination dynamics and the time to the epidemic peak? You can't just re-scale the y-axis on your plots. And now if you want to start with 500 cases in the real population of 5,000,000 but it is implemented as 1 case in an actual population of 10,000 people, then you will get tons of stochasticity when there shouldn't be with 500 initial cases.

[ckerr-IDM]

@dlchao it's a good point, we'll need to be able to scale down as well as up! but the idea is to get around exactly that -- we start with 500 cases in a population of 50k, and then when we get to 5000 cases we rescale to 500k, etc. -- so we never have fewer than 500 simulated cases but also never have more than 5000

[inc0]

Another thing is that we're rescaling by factor of 2 by default and scaling multiple times (every few days once epidemy picks up).

example: 10k internal population and 2M target population, start with 10 infections:

• keep going until you reach 5000 or less susceptibles
• cut make half of non-susceptibles susceptible again, multiply all numbers by 2
• reach 10000 (5000 * 2), multiply by 2 again
• repeat

[ckerr-IDM]

as i mentioned in my comment on the PR we definitely don't want to let prevalence reach 50% -- maybe not even 5% -- before rescaling

[inc0]

I've set it up to 5% and it's configurable, we can experiment with even lower thresholds if need be

[dlchao]

How do you make transmission trees when you have dynamic rescaling of the population?

[ckerr-IDM]

you wouldn't be able to use transmission trees with this -- assuming you probably wouldn't be doing a full transmission tree for >1m infections